fix(webv2): let the browser-driving scripts run on Windows

[Pfannkuchensack]

Summary

Three scripts drive a real browser against a built app: the architecture performance measurement,
the accessibility journeys and the project-file journey. Each starts pnpm exec vite preview and
tears it down with process.kill(-pid, ...). Neither half works on Windows, so all three fail
before a browser opens:

Error: spawn pnpm ENOENT

pnpm on Windows is a .cmd shim, and Node refuses to exec one without shell: true.

shell: true on its own would trade that for a worse failure, which is why this does something
else. It puts cmd.exe between the script and vite, so preview.pid is the shell's — teardown
signals the shell and leaves vite holding the port. The next run then fails on --strictPort, or,
if it connects anyway, measures against whatever build the orphan is still serving. That is not
hypothetical: it happened while developing this, and produced a confusing chunk-hash mismatch two
runs later.

So this avoids the shell entirely. scripts/preview-server.mjs starts vite's own entry under the
same Node — no shim, no wrapper, preview.pid is the server — and gives teardown a Windows branch,
since there are no process groups there: taskkill /T walks the tree, and its "no such process"
status is mapped to ESRCH, which run-project-file-journey.mjs relies on to tell an
already-exited server from a real failure.

Linux and macOS keep exactly the behaviour they had: detached: true, group kill, same signals.

Related Issues / Discussions

Found while re-measuring browser-baseline.json for #146, where the performance gate could not be
checked locally at all.

QA Instructions

On Windows, from invokeai/frontend/webv2:

pnpm test:performance:build && pnpm test:performance:browser
pnpm test:project-files

Both pass here and leave nothing listening on 4176/4180 afterwards; on main both stop at
spawn pnpm ENOENT. pnpm test:fixtures covers the new module (26 tests).

On Linux and macOS the same three scripts should behave exactly as before — that is the half worth
a reviewer's eye, since this PR cannot be verified on Windows and Linux by the same person easily.

pnpm test:accessibility:browser now starts on Windows and gets as far as the topbar journey,
where it times out waiting for an Invoke from button. That is a pre-existing failure this makes
visible for the first time rather than a regression — the script could not run on Windows at all
before, and CI runs check:architecture, which does not include it. Not addressed here.

Review

Agent-assisted. Independent review subagents were not run for this one; it is small and was
verified by execution rather than inspection. The taskkill status-to-ESRCH mapping is the part
most worth a second opinion — it is hand-written, and preview-server.test.mjs pins it (the test
was made to fail on purpose by changing the status code before being trusted).

Compatibility / Rollout

No product code. Developer tooling only; CI runs on Linux and is unaffected.

Checklist

• The PR has a short but descriptive title, suitable for a changelog
• Meaningful regression coverage added / updated where needed; obsolete tests/code removed
• Persisted-state and API changes include required migrations / compatibility validation
• Relevant performance/efficiency opportunities considered; material claims have evidence
• Material review findings resolved and relevant checks rerun — self-review only, see Review
• Documentation added / updated (if applicable)
• Updated What's New copy (if doing a release after this PR)